Myocardial infarction detection based on deep neural network on imbalanced data
Myocardial infarction detection based on deep neural network on imbalanced data
复制标题
DOI：
10.1007/s00530-020-00728-8
复制
发表时间：
2021-01-06
影响因子：
3.9
通讯作者：
Abd El-Latif, Ahmed A.
中科院分区：
文献类型：
--
作者：
Hammad, Mohamed;Alkinani, Monagi H.;Abd El-Latif, Ahmed A.
心肌梗死（MI）是流向心脏的血液的急性中断，其导致心脏遭受血液缺乏和缺血，因此心肌受损，细胞可以死亡并失去功能。尽管心肌梗死在世界范围内的发病率很低，但它仍然是一种常见的导致死亡的疾病。因此，早期检测MI信号可以降低死亡率。提出了一种基于深度卷积神经网络（CNN）的MI自动检测方法。提出的CNN是一种端到端模型，不需要任何机器学习阶段，只需要一个阶段来从输入信号中检测MI。在数据不平衡的情况下，我们使用一个名为焦点损失的新损失函数来优化我们的深度模型，通过将损失间接地构成那些困难类别中的焦点来处理这种情况。在验证中使用了Physikalisch-Technische Bundesanstalt（PTB）数据集，将信号分类为正常和MI。我们的技术与该领域最先进的技术一起表现出平均准确性和F1得分的增加。结果表明，焦损失提高了9%的检测精度检测MI信号。总之，所提出的方法实现了98.84%，98.31%，97.92%和97.63，分别使用焦点损失和89.72%的总准确率，88.52%的精确度，81.11%的召回率和83.02%的F1分数的总体准确率，精确度，F1分数和召回。我们的方法使用局灶性损失是一个有效的工具，执行一个快速和可靠的MI诊断，以协助心脏病学家在早期发现MI。
Myocardial infarction (MI) is an acute interruption of blood flow to the heart, which causes the heart to suffer from a deficiency of blood and ischemia, so the heart muscle is damaged, and cells can die and lose their function. Despite the low incidence of MI in the world, it is still a common disease-causing death. Therefore, detecting the MI signals early can reduce mortality. This paper presented a method based on a deep convolutional neural network (CNN) for the detection of MI automatically. The proposed CNN is an end-to-end model without requiring any stages of machine learning and requires only one stage to detect MI from the input signals. In the case of imbalanced data, we optimize our deep model with a new loss function named the focal loss to deal with this case by constituting the loss indirectly the focus in those difficult classes. The Physikalisch-Technische Bundesanstalt (PTB) dataset was employed in the validation to classify the signals to normal and MI. The performance of our technique alongside state-of-the-art in the area shows an increase in terms of average accuracy and F1 score. Results show that focal loss improves the detection accuracy by 9% for detecting MI signals. In summary, the proposed method achieved an overall accuracy, precision, F1 score, and recall of 98.84%, 98.31%, 97.92%, and 97.63, respectively using focal loss and overall accuracy of 89.72%, a precision of 88.52%, a recall of 81.11% and F1 score of 83.02% without using focal loss. Our method using focal loss is an effective tool to perform a fast and reliable MI diagnosis to assist the cardiologists in detecting MI early.